Massively parallel DNA sequencing platforms have recently become broadly available (see, for example, Mardis, E. R., “The Impact of Next-Generation Sequencing Technology on Genetics,” Trends Genet. 24:133-141 (2008), and Wold, B., et al., “Sequence Census Methods for Functional Genomics,” Nat. Methods 5:19-21 (2008)). Several platforms operate at a fraction of the per-base costs of conventional electrophoretic sequencing, but produce sequence reads that are over an order of magnitude shorter and less accurate. These short reads have information content such that most are uniquely mappable to genomes with an existing reference assembly, enabling a variety of “sequence census” applications (see Wold, B. and Myers, R. M., “Sequence Census Methods for Functional Genomics,” Nat. Methods 5:19-2.1 (2008)). However, the short lengths and high error rates impose significant limitations on the utility of short reads for applications such as de novo genome assembly, full length cDNA sequencing, metagenomics, and the interrogation of non-unique subsequences of assembled genomes. Towards addressing these limitations, this invention provides methods and compositions that enable the clustering of short reads derived from the same kilobase-scale fragments. Each cluster of short reads can then be locally assembled in silico into a single long read or a mate-pair of long reads, which are referred to as “subassemblies,”